⚠ Good-Enough & DepreciationModerate threat

The upgrade treadmill assumes that last year's hardware becomes a competitive liability, and the danger is that this assumption weakens as the technology and the workloads mature. If older Nvidia chips remain perfectly useful — especially for inference, which is less demanding than frontier training — then customers can stretch their upgrade cycles, buy prior-generation hardware, and slow the frantic re-buying on which so much of Nvidia's growth depends.

The threat is amplified by the vast installed base Nvidia has already sold. Every powerful chip shipped in the boom becomes, in time, a source of cheap secondhand compute and a reason not to buy new, and as the pace of model improvement eventually slows, the marginal benefit of each new generation shrinks relative to its cost. A treadmill running on the fear of falling behind loses power the moment behind stops being so costly.

What sustains the treadmill for now is that the frontier is still moving fast, efficiency gains are large enough that new chips genuinely lower the total cost of a workload, and the biggest buyers still compete ferociously to train the best models. As long as the race is hot and each generation is markedly better, the incentive to re-buy holds, and Nvidia keeps engineering reasons to upgrade.

On balance, moderate. The good-enough dynamic and a growing installed base of durable hardware are real long-run brakes on the re-buying, and they will matter more as workloads mature and the frontier's pace eventually eases — but while the race stays frantic and each generation delivers large, benchmarked efficiency gains¹, the treadmill keeps turning and the premium keeps recurring.